Probabilistisch-logische Inferenz relationaler Situationsbeschreibungen aus Verkehrsbildfolgen

In modern automotive systems, driver assistance based on machine perception is increasingly gaining importance. This coincides with a steady rise in automation concerning the assistance function. As a result, the demand on behalf of the complexity of the traffic scenarios to be handled grows. Vehicles capable of avoiding accidents, do not only have to be able to percept their environment with sensors, but to analyze and interprete the current situation in terms of their own actions. A knowledge-based description of the traffic situation, including all occurring objects, their attributes and interrelations is therefore essential for scene interpretation and successive actions that are safe and reasonable. This thesis presents a model based on Markov Logic Networks. It is used for probabilistic-logical based inference of a situation description enhanced with probabilities given variable traffic scenes. For each segmented object hypothesis, class and relations to other object are estimated within a single inference step. The probabilistic-logical formalism Markov Logic Networks allows for an expressive relational representation of complex domains. In this representation it is possible to model uncertainties and conduct probabilistic reasoning, which is exemplarily evaluated on processed video sequences.